Without limiting the scope of the present invention, this background of the present invention is described in connection with connection-oriented communication protocols, such as the Transmission Control Protocol (“TCP”).
There are several software-based techniques, such as active replication, semi-active replication, primary-backup, and rollback recovery, that attempt to provide computational fault tolerance. These techniques focus on protecting the state of applications residing within user-level processes. Efforts to apply these techniques to protect the state maintained within the operating system, however, have been largely unsuccessful. Typically, these efforts assume very restricted interactions with the operating system and interactions with the environment through restricted interfaces. In practice, however, applications interact with the operating systems in unrestricted ways and communicate both with the full set of devices on the local machine and with other processes through connection-oriented communication protocols such as TCP. As a result, even applications that rely on sophisticated techniques for recovering from crashes at the user process level do not achieve the desired level of fault-tolerance.
For example, a fundamental problem in computer networks is to determine an optimal path from a source to any destination in the network. This problem is especially critical in networks of the size of the Internet, where computer and network failures constantly modify the topology of the network. To monitor the network, a set of dedicated IP routers and wide-area network core switches (for Frame Relay and/or ATM) run special protocols, such as the Border Gateway Protocol (“BGP”), which they use to exchange information with their peers when they detect a change in the topology of the underlying network. To achieve greater reliability, many IP routers use a primary-backup fault-tolerant protocol implemented on a hardware-supported process-pair architecture. When a primary router process in this architecture fails, the backup process automatically begins to function as the router and the state of the backup process will typically be indistinguishable from the state of the primary process at the time it failed. Computer systems that implement this process-pair architectures are significantly more expensive than their non-fault tolerant counterparts.
Unfortunately, even these expensive dedicated computer systems cannot prevent the failure of the primary process from having undesirable side-effects. The reason is that these Systems are unable to mask the loss of application state at the operating system level, and Border Gateway protocols are built on top of TCP, whose state is maintained within the operating system. When the primary router fails, the TCP connections that it was maintaining with all the peers participating in the BGP are severed. The surviving peers interpret the loss of these connections as a failure of the primary router and initiate state changes through the BGP to reestablish routing around the failed primary router. When the backup router takes over for the primary router, it reestablishes BGP sessions with its peers and routing can resume through the recovered component. Before the take-over completes, however, the primary router has incurred costs in network capacity, delays, lost packets in transit, etc. Even though these IP routers and core switches include hardware redundancy to tolerate failures and have the ability to fail over the application services in less than a second to a backup process, the architecture of the BGPs still exhibits the side effects of the severed TCP connections, and the rest of the network still transitions the faulty IP router to an out of service state.
This example demonstrates a more general problem that occurs whenever applications running on top of connection-oriented protocols, such as TCP, use the loss of a connection as a failure-detection mechanism. As a result, the application's response to the loss of the connection may generate unwanted side effects, such as a change in the content of the network routing tables, even if the failed node has state-of-the-art fault-tolerance capabilities.